NVIDIA DLSS 5 Technical Preview Review – Neural Rendering is Here

Introduction

We first heard about DLSS 5 in March, when NVIDIA unveiled it at GTC and called it the biggest change to real-time graphics since ray tracing. The reaction was not what the company had hoped for—people were concerned that their games would no longer look correct and artistic intent was lost. A few weeks ago at SIGGRAPH, NVIDIA showed a second update with substantial improvements, clearly having listened to the negative feedback from the community. Last week at Gamescom they went further again, and it is now official: DLSS 5 launches with NBA 2K27, which goes live on Friday. It is supported on all GeForce RTX 50-series GPUs, on RTX 50 laptops, and on GeForce NOW.

In this article we will work through NVIDIA’s technical briefing from Gamescom, presented by Edward Liu, who leads research across all DLSS technologies, and Gabriele Leone, director of the content technology team—the group that builds the developer-facing tools. Neither is from marketing, and Edward opened by saying so himself. Both halves were considerably more technical than a typical launch presentation, which is what makes them worth going through in detail. We also had the opportunity to ask questions, and we took it; those are placed inline throughout the article wherever they are relevant.

DLSS 5 is fundamentally different from every DLSS technology that came before it. Upscaling and frame generation are reconstruction problems: the renderer was trying to draw something specific, and the model’s job is to work out what that something was. There is a right answer, and you can measure how close the model gets to it, and train an AI towards that goal.

Neural rendering has no such answer. DLSS 5 adds material on top of the game’s output—photorealistic lighting, subsurface scattering through skin, contact shadows, the sheen on a piece of cloth—and for that there is no ground truth, no “100% correct” frame. That makes it much harder to verify, and it opens the door to ambiguity in a way none of the other DLSS technologies did.

NVIDIA’s approach to that problem is to let people hand-guide what the model does. They have trained it so that it can add lighting and material response on top of a frame, but cannot touch the geometry underneath—the silhouettes, the surface normals, the shapes the artists actually built stay exactly as the renderer drew them. It is a clever approach, and it rules out the failure everyone was worried about back in March. On top of that, there is now a set of controls exposed to developers that lets them dial in how much of the effect they want, apply different amounts to different parts of the image, or switch DLSS 5 off completely for the parts they would rather leave untouched.

That means the useful question about DLSS 5 is not “is it accurate,” because no such thing exists in this scenario. What can be asked is “does it look good, and does it still look like the game the artists shipped?” Here is NVIDIA’s Zorah demo with it off and on—a good place to start forming your own opinion.

Compatibility

This is the part that will disappoint a lot of people. Asked directly whether DLSS 5 is RTX 50-series only, NVIDIA answered “Yes, it is.” It runs on Blackwell’s fifth-generation Tensor cores, and there is no fallback path for RTX 40-series or anything older.

Asked separately whether older RTX cards might get the option to run it anyway, even badly—the way Super Resolution and Ray Reconstruction are available all the way back to Turing—the answer was a non-answer: “DLSS 5 is the largest, most complex model NVIDIA has built, and their priority has been optimizing it for RTX 50.” That is not “never,” but it is not “we have a plan for legacy” either.

It also matches what we found when we pulled apart the leaked DLL last week: executable code for Blackwell only, behind an explicit runtime gate reporting an unsupported architecture. At the time we cautioned that a pre-release build was not proof of the actual plan for public release. It turns out it was. Worth noting alongside is that modders have already forced the leaked DLL to run on RTX 40-series, so whatever that gate is, it is not that Ada physically cannot execute the model. There is a great deal more to that story, more on that on the next page.